Getting pretty off-topic: database systems

[James Wilkinson]

Vincent Trouilliez wrote:
> The only ones I ever saw, was last summer, when I worked for a month at
> Bull, assembling and configuring their machines, well, learning with a
> technician teaching me really, basically IBM H/W with AIX. I was
> learning on their smallest offering, a 42U (IIRC, was about 2meter high)
> cabinet, a couple 8 CPUs nodes (PL820R anyone ?), each with 6 drives
> IIRC, and of course 2 or 3 disk arrays/DAS (can't remember... 14 disks
> each ?) and a few more similarly sized cabinets with no nodes, just
> fully loaded with disks.

Interesting. I know that IBM offers smaller p-series servers: I'm
working on them. (There's supposed to be little difference between
Bull's AIX servers and the IBM ones).

> Can't remember the size, I think it was 73GB
> 15K rpm drives. So rougly 100 disks or so per cabinet, 7.5 TB just for
> one little cabinet ! :-O  We used RAID5, so I don't think we lost much
> capacity, compared to RAID1 anyway...

Actually, RAID 5 *isn't* recommended for most databases. These days,
most databases have more of a requirement for speed than disk capacity:
you can get disks big enough, but you can't get disks that spin fast
enough. In that case, you measure disk performance by access time and
the number of "spindles". So RAID 1's halving of capacity isn't an
issue. It only takes two writes (on different disks) to update it.

But with RAID 5, when you're doing lots of updating of small 8K blocks
(that's a lot less than stripe size), you need to read enough data to be
able to work out what the parity data should be. And that usually means
at least two reads followed by two writes one disk revolution (~0.006
seconds) later. Much slower.

But I suspect your customers would then re-configur the disks the way
they wanted, anyway.

> Nice machines (if just a tad noisy...), too bad I stayed only 4 weeks
> there, hardly enough time to learn anything properly really :o(
> Would you believe it, instead of AIX, they also offered Mandrake on
> these monsters, upon explicit customer request of course.
> Would be fun to see Ubuntu run these things, but does Ubuntu support/run
> on IBM servers ? 

It's not beyond the bounds of possibility: they aren't too dissimilar to
Apple Macs.

> But about fragmentation on these, no idea... can't even remember the
> file system used. I just remember that we tested the disk of the nodes
> with 'dd', then created a few RAID 5 devices to check that the DAS were
> working, and that's about it... ready for 24 hours of stress-testing. 

If it's AIX, it would be JFS or JFS2 (what the Linux world knows as
JFS). And no, it's not particularly prone to fragmentation.

> I wonder what the performance of the DAS was... if only I knew how to
> test it back then ! Too late now :o(

Transactions per minute. Either some standardised version (TPM) or
(better) the number of transactions while doing typical work for the
company.

James.

-- 
E-mail address: james | Let He who Taketh the Plunge
@westexe.demon.co.uk  | Remember to Return it by Tuesday.